Mr Wisp
patreon
The Rest: November 2023
🕑 Added 2023-11-01 04:00:04 +0000 UTCHello! Before you proceed, please take the time to note the THREE folders containing a combined 400+ AI-generated images attached to the end of this post! There are three folders this time because the average image size has increased significantly, which pushed the zip folder size past what Patreon could handle for an individual file. These images come from three separate sessions, each of which I'll describe individually, in chronological order. I'll also be including the images I used as the basis for each session to give you all a better sense of what NovelAI had to work with. For some basic notes on my AI image generation process as well as my intent/goals with the Re-AImagined content, check out my previous 'The Rest' posts, where I go into those things in detail.
Lani Session #1 - Late August 2023
This was a very simple and straightforward session, and I don't have a ton to say about it specifically, so I'll use this space to talk about the general experience of trying to reproduce Lani in NovelAI. My goal with all the Lani sessions I've done has been to illustrate a vague weight gain progression starting with fit, muscular Lani and ending with a size reflective of Lani's max weight in Axugaem2 (you'll have to wait for another Lani gallery to see if I was able to get there, as there were just too much Lani to fit into one month's gallery!) Thus, session #1 is mostly fit Lani, whereas session #3 is entirely fat Lani, with session #2 marking a transition between the two.
Reproducing Lani herself in NovelAI ended up being a similarly simple and straightforward process overall. This was somewhat of a surprise to me for one reason: I didn't think NovelAI would be able to make sense of her hair. But as I realized upon doing some research into hairstyles, I am stupid. I originally adapted Lani's hairstyle from a sheet of hairstyles I found while browsing DeviantArt, and I never bothered to look into it any more than that. But it turns out that Lani's hairstyle is basically just something called an 'inverted bob' (albeit one that follows the Joey Wheeler school of hair physics.) Adding that tag and a few others to do with hair length and bangs was all it took to start producing results that looked like Lani. Sure, the hair wasn't as big or M-shaped, but those traits were more a result of me not knowing how the hairstyle I was drawing worked for several years, so I don't consider them essential to Lani's character.
Once I got the hair sorted, the rest of Lani was about as simple as it gets in terms of tags. There were the usual inconsistencies with the colour of her shirt (and sometimes headband), but as long as the result was something I could imagine Lani wearing, I didn't sweat over it. Her key necklace ended up being well-represented in most of the results I generated, so much so that I wasn't bothered about the results where it was more ambiguous in shape. A couple of results made the key into a cross, which by coincidence is lore-friendly to air worshippers in Imperium (the cross represents the four winds, hee hee), so that was a nice surprise. One area where NovelAI excelled was with Lani's musculature, which is always nice because that isn't well-represented in any art of Lani from before two years ago (including the reference of her that's used in some of her Axu2 graphics, and which I've attached with the zipped folders below.) Muscles aren't my forte, but when I finally draw up a new Lani reference, I'll be taking a cue from novelAI's take on her. I'd even consider some of the results from this first session to be more accurate depictions of Lani than my own art, which is definitely a first for these galleries!
The background for this session is meant to represent the plains outside Flamebane, Lani's home village. This is a high-altitude area without very many trees, and only a few taller mountains surrounding it. Thus, I was fully satisfied with what NovelAI came up with, even if it was fairly basic. The 'depth of field' and 'vanishing point' tags helped give the results the sense of wide, open space I was looking for as well.
As far as the various settings go, I went straight to 0.99 strength for all of these sessions, though I did spend a bit of time playing around between 0.9 and 0.95 as well, mostly to see how that affected the chubbier versions of Lani. What fiddling I did do with the various other sliders was almost entirely for the later sessions, as novelAI knew what it was doing for the majority of session #1.
Lani Session #2 - Late August 2023
This session is really two smaller sessions that I chose to group together due to their similarities. The images of Lani standing outside at sunset were done first, as a direct followup to session #1. The images of Lani in the dojo (no prizes for guessing whose dojo) were done a little later and were mostly just for the sake of putting her in a new and slightly more interesting setting. Insert all the stereotypes about inner fire and fiery passion and such here.
NovelAI handled the musclegut content pretty well here. I went over this a bit in the notes for the Raziel sessions, so I won't belabour the point, but the 'big belly' and 'muscular female' tags seem to play relatively nicely with eachother as long as the former isn't overemphasized. Other tags, like 'warm colours', [[bloom]], 'backlighting', and 'bokeh' helped especially to emphasize the fiery heat in the dojo images. Finally, the 'undersized clothes' tag is one that I used from the dojo session onward, as it helped create some awesome belly/shorts interactions, which are fully a weakness of mine!
In terms of tags that didn't work so well, I tried working with the 'underboob' and 'sideboob' tags during these sessions as a way to further emphasize how much Lani had grown relative to her clothes, and to give a sense of uh, dynamism to her fighting moves? (I wanted to see underboob. I have no excuse.) But the results weren't really what I was looking for. By that I mean, even when de-emphasized, these tags warped the rest of the image to prioritize there being under/sideboob, even if all of it ended up looking awkward and unflattering. In cases where that didn't quite happen, I still felt like the tags weren't particularly helpful, as they had a tendency to change the focus of the image from 'Lani is a cute and thicc brawler' to 'LoOk At LaNi'S bOoBs!!' Since Lani isn't particularly fond of or interested in showing off her chest, the vibe just didn't feel right, even with the fact of her clothes being undersized. If I redid this session, I'd remove these two tags from the prompt, but as it is I left them in longer than I should have. With all of that said, they did lead to a couple of believable results involving Lani 'airing out' her top, which I wouldn't have gotten without them, so perhaps I'm being too critical.
Lani Session #3 - Mid October 2023
This is my favourite out of all the Lani sessions I've done to this point, including those not yet posted here. Is that because the subject matter is a very plump Lani sitting on a bed, with the implication being that she just ate a shitton of pie? Yes. Yes, it is. But it's also because the overall quality of this session is among the best I've seen across all the sessions I've done. There's a reason why this session makes up the bulk (ay lmao) of this month's gallery: there were just too many results that I wanted to include, and none of them felt as 'same-y' as those from sessions #1-2. I'm a fan of putting characters on beds for practical reasons (no, seriously), as it eliminates a lot of the issues that come with the 'sitting' pose in AI art, such as having to deal with chairs that look like they were designed by M.C. Escher, for instance. The perspective is usually inoffensive, the framing can be more intimate (in more than one sense of the word), and in case I didn't imply this clearly enough, it's really, really hard to make a bed look janky, at least from my experience. That said, the reason I don't just put all of my characters in beds and call it a day is that doing so would be too easy. And it'd be boring. And perhaps most importantly, not every character of mine spends a lot of time in bed, especially not eating in bed. True Fact: Eating in bed increases your chances of getting insects, and getting insects increases your chances of getting spiders, so you know by that logic that there ain't no chance Nell's eating anything in bed. Lani is normally just as careful, but as is her axiom, all bets are off whenever pie is involved. Even when pie isn't involved (which is rare), she tends to spend a lot of downtime in her room at Gusbuster Base, since the alternative is risking an interaction with Linda, which never ends well for Lani. So in this case, the stars aligned and putting Lani on a bed ended up being the right move not just practically, but lorewise too!
I also think the quality of Lani's fat here is exceptional by nearly any standard. This definitely owes to the simplicity of the setting and pose, but I also suspect that moving the fat-related tags to the front of the prompt helped as well. This is a technique that some folks on Reddit claim increases the quality of the results, at least as pertains to the tags you've chosen to prioritize. This session was one of my first experiments with that method, and based on the results I'm inclined to believe it works. Obviously, future sessions will make it clearer to me if and how much this method really affects the quality of the results, but seeing as I didn't do much else with the tags themselves that was different from previous sessions, for now I'm optimistic!
Standout tags from this session include 'warm lighting', 'orange theme', 'dynamic angle', 'cinematic angle', and 'cinematic lighting'. These tags all helped bring the results together and give them the sort of oomph I was looking for. 'Orange theme' specifically was chosen since it was the best midpoint I could think of between yellow and brown (Lani's two main colours), and orange is a nice, warm colour for a cozy day spent eating pie in bed. Of course, including this tag turned a few of Lani's clothing items orange at times, but I ended up liking a lot of those results. 'Lani could rock orange shorts' is the big takeaway here.
A few things I did differently with the settings this time around: for one thing, I've been starting with the steps set to 50 rather than 28, which means roughly that NovelAI takes more steps to consolidate the details of each image generation. This can sometimes result in more finished looking images, but it can also lead to an overdone quality that ends up looking worse than if the steps were kept at a lower value. Again, from my reddit-browsing, I found a few accounts that spoke positively about using a higher number of steps, so I figured I'd give it a try. I'll continue playing with this technique in future sessions, as I'm curious just how impactful it actually was on this session. I can at least say that it didn't hurt the results, so that's encouraging!
I also kept the prompt guidance value low throughout this session (between 7.5 and 8), which seemed to gel with the rest of the settings. Granted, setting it to a higher value would almost certainly have helped there be more actual pie in the images (something I was hoping for but never got), but as I'll note in a paragraph or two, the 'pie' tag had its own issues...
Now for the parts of this session that didn't go so smoothly: Lani's shorts and the inclusion of pie as a prop. The specific problem with Lani's shorts this session was getting them to be 'unbuttoned' or 'open shorts' (both tags I emphasized the crap out of in an attempt to make them work). I'm not sure what I did wrong here, but NovelAI was insistent throughout this entire session and every subsequent Lani session that Lani's shorts not only be done-up, but devoid of buttons as well. This persisted even after I put the 'gym shorts' tag at the top of the undesired content list, and there were precious few exceptions. My hope was to generate some images where Lani's belly was spilling out over her undone shorts rather than just straining against them, which is admittedly one hell of a first world problem, but it still would have been nice to see. That said, the sheer quality of Lani's belly throughout this session more than made up for the lack of open shorts, and there were plenty of results that gave me more or less what I was looking for but with the shorts still done up, so I certainly can't complain. Still, I'm going to work more on figuring out how best to tell NovelAI to unbutton the damn shorts in future sessions.
Now, let's talk about pie. You didn't think we'd get through an entire document on Lani without discussing pie, did you? I didn't go as hard in this session as I did with subsequent sessions, but this was still the first indicator to me that pie is annoyingly hard to make appear consistently using NovelAI, at least with the settings I was using. 'pie slice' seems to be the most reliable tag to use if you want there to be pie in your images, as nearly every result from this session that featured pie featured it solely sliced. 'Eating pie' also helped, I'm sure. Now, it is true that not only did I not emphasize these tags very much, but I also had the prompt guidance value set low, which means that NovelAI wasn't basing its image generations off the concept of 'pie' as directly as it could have been. The reason I didn't change either of these things has to do with a fundamental issue with the 'pie' tag itself. You know how NovelAI's anime model was trained using Danbooru's image database? Yeah, y'know how Danbooru's image database includes a significant amount of hentai? Turns out there's usually only one type of pie that appears in hentai, and it is perhaps the only kind that Lani doesn't want anywhere near her mouth. See images 'ln148 2' and 'ln152 4' in the attached folders for two very, very tame hints at what I'm talking about. This unwanted duality of 'pie' is probably the biggest problem I'd like to resolve in future Lani sessions, as Lani without pie is a sad Lani, but Lani with that kind of pie is a very, very angry Lani, and that's much worse.
